New Delhi: Meta (formerly Facebook) is building a first-generation custom silicon chip to run artificial intelligence (AI) models, and as it breaks new ground in AI research, AI computing needs will grow in the next 10 years. He said it would increase dramatically over the years.
An in-house custom accelerator chip called MTIA (Meta Training and Inference Accelerator) provides greater computational power and efficiency than CPUs and is tailored for internal workloads.
“By deploying both MTIA chips and GPUs, we are able to improve performance, reduce latency and increase efficiency for each workload,” said Santosh Janardhan, vice president and head of infrastructure at Meta. increase.
The company is also planning a new AI-optimized data center design and a second phase of a 16,000 GPU supercomputer for AI research.
“These efforts and additional ongoing projects will enable us to develop larger and more sophisticated AI models and deploy them efficiently at scale,” added Janardhan.
The next-generation data center will be an AI-optimized design, supporting water-cooled AI hardware and a high-performance AI network connecting thousands of AI chips for data center-scale AI training clusters.
“It is also faster and more cost-effective to build, complementing other new hardware such as the MSVP (Meta Scalable Video Processor), our first in-house developed ASIC solution designed to power the ever-growing video market. Meta’s workload,” Janardhan said.
Meta’s Research SuperCluster (RSC) AI supercomputer, which the company believes to be one of the fastest AI supercomputers in the world, is the next generation of large-scale AI to power new augmented reality tools, content understanding systems and real-time translation. Built to train models. such as technology.
It has 16,000 GPUs, all accessible across a three-level Clos network fabric that provides full bandwidth to each of the 2,000 training systems.
“By custom-designing much of our infrastructure, we are able to optimize the end-to-end experience from the physical layer to the virtual layer to the software layer to the actual user experience,” said Mehta. Stated.
